3-2. Assessment of the Sentimental Analysis Tools

There are several possible causes for the accuracy differences between the tools. For instance, “domain-transfer problem” would be one of them. Sentimental classifiers like “Corenlp” have an ability that improve themselves. Supervises such as human programmers can teach them the ways to deal with specific domains. By doing so, those software is able to show high performance on the supervised field. On the other hand, if the supervised sentiment classifiers are used to analyze the different domains (unsupervised domains), their performance becomes very low. This phenomenon is called "domain-transfer problem" (Tan, Cheng, Wang, & Xu, 2009). Probably, this explains the cause of the lowest accuracy of “Corenlp” among the three applications. Unlike other two which were Web-based software, “Corenlp” was Java-based program and must had been trained for the domain of “News headline” in order to evaluate it properly. The other reason for the “Negative-oriented” analysis of Sentimental classifiers would be the latent nature of “News Headlines”. As some people suggested,“News Headlines” tend to be negative because readers are more attracted by negative headlines than positive ones (Headlines: When, 2013). What is more, unfortunately, there is a high possibility that the immatureness of the research in terms of the survey process caused the inaccurate outcome.

(All the results for this investigation are appended to the end of this thesis as appendix.)